Data-Driven Small-Disturbance Stability Assessment and Preventive Control in Mixed AC/DC Low Inertia Power Systems

Abstract

Due to a growing number of converter-interfaced sources in AC grids, the non-linear characteristics of power systems have been continuously increased. The highly non-linear power systems pose a challenge to a conventional small-disturbance stability analysis, modal analysis, since this technique requires a linearised model of power systems in the calculation. Therefore, small-disturbance stability assessment needs to be completed and the poorly damped inter-area mode needs to be improved as quickly as possible. The reduced-order GB power system representing the operation in 2030 is developed as a test system used throughout this thesis. This system is suitable for study the small-disturbance stability in the mixed AC/DC system since it has two groups of synchronous generators oscillating against each other, representing an inter-area mode, and it contains a large number of wind generators and HVDC connections. In the modelling of complex mixed AC/DC system, modelling a large number of dynamic HVDC systems require a high level of modelling skills and computational resources. In this thesis, a new index, mode shape quantification, is developed to identify the locations where the dynamic HVDC model can be replaced by the simplified HVDC model without a negative impact on the small-disturbance stability results. Implementing the dynamic HVDC model only at locations, specified by the developed index, can significantly reduce the modelling complexity in the large power system without causing errors in the small-disturbance stability results. In past literatures, the influencing variables on the damping of the inter-area oscillation were investigated in the system with only small numbers of renewable sources. This thesis further analyses the influencing variables in the system with high penetration of converter-interfaced sources and variations in network structures. The analysis of the influencing variables on the damping of the inter-area oscillation is implemented under single- and multi-dimensional variations. The correlation analysis in single-dimensional variations shows that there are underlying relationships between some system variables and the damping of the inter-area mode. However, the strength of these relationships decreases significantly in the multi-dimensional variations. To provide accurate real-time prediction of damping ratios of an inter-area mode, machine learning method, XGBoost algorithm, is implemented to form the relationship between the variables, directly simulated from the test system, and the damping ratios of the inter-area oscillation. The XGBoost model is firstly developed to provide the real-time damping prediction of the fully-connected system. The two-level learning approach is developed in this thesis to provide damping prediction of the power system under different network topologies. In the proposed technique, the network topologies with similar damping characteristics are grouped using the hierarchical clustering model, and the XGBoost model is used to predict the damping ratios of each cluster. The ML-based prediction model is used in conjunction with Bayesian optimisation to perform fully data-driven generation rescheduling. The data between XGBoost model and Bayesian optimisation are iteratively transferred until the optimal power patterns are found. The proposed rescheduling technique is completed without any use of a linearised model of power system.

Date of Award	1 Aug 2021
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Robin Preece (Supervisor) & Ognjen Marjanovic (Supervisor)